What an internal AI policy must cover

Approved tools and access roles

The policy begins with a binding list of tools employees may use for work tasks. Group them by purpose: text generation, code generation, image generation, data analysis, translation. For each tool state whether it is a corporate licence such as Microsoft 365 Copilot or ChatGPT Enterprise, or a permitted free version with data restrictions. Assign roles: who may access paid models, who may use only local models, who has no access at all. For example, a marketing team uses Midjourney Enterprise for concepts, developers use GitHub Copilot Business, while HR may not use any generative model for processing CVs.

Data classification and input rules

Define data sensitivity classes and, for each class, permit or forbid sending data to AI. Typical classes: public data may go anywhere; internal data only to corporate tenants with a confidentiality agreement; confidential or personal data nowhere into the cloud. Explicitly forbid pasting email correspondence, contracts, payment card details, health data or source code of critical systems into public interfaces such as ChatGPT Free, Gemini web or Perplexity. Note technical controls: DLP rules that block insertion of national identification numbers into browsers, browser extensions that warn when data is entered into AI fields.

Explicit prohibitions and risk scenarios

List concrete actions that are prohibited without exception. These include automated decisions about employees such as hiring, termination or bonuses based solely on AI output; generation of legal documents without lawyer review; publishing AI-generated text as expert articles without labelling; using AI to bypass security controls such as writing exploits or generating phishing emails for tests without authorisation. State sanctions ranging from a warning to termination of employment in repeated cases.

Output verification and human oversight

Require verification of every output before production use or client delivery. For text: check facts, sources, logic, style. For code: static analysis, unit tests, peer code review. For data: confirm the model has not hallucinated numbers in recalculations. Institute a human-in-the-loop rule: the responsible person signs off the output by name. Example: an analyst uses Copilot for a SQL query, then verifies the query plan and tests on a data sample before production execution.

Accountability, logging and policy review

Assign responsibility for damage caused by AI error to the employee who approved and used the output, never to the model provider. Mandate logging of cases where AI materially influenced the result, recording prompt, output and decision. The policy is signed by the CISO or CDO, valid for 12 months with mandatory review at six months or after legislative change such as the AI Act, a model provider change, or an incident. The review team comprises security, legal, IT, HR and the business owner.

Living process not static document

A policy is not a one-off PDF filed away but a living process. Tools change, models update, legislation sharpens. Without regular review and assigned personal accountability the document becomes obsolete on the day it is issued.